Input and Output - DMA I/O

Multi-Bus Architectures

• Multiple Processors

  - One way to increase the speed at which a process is performed is to increase the number of processors working on the problem. Furthermore, by dividing a task into subtasks, individual processors can specialize; specialization means that processes can be performed by those most effecient at their performance. This is analagous to the mass production factory techniques.

• Sharing a Common Memory

  - Multiple processors each with their own external bus, can be connected to a shared memory. This memory may be used as a media for communication of data between processors as well as for the individual processor's needs. A processor's memory may all be shared with some other processor, or a processor may share only a portion of its memory. When a portion of a processor's memory is available and shared with another processor, that portion of memory is said to be "mapped" to the other processor.

• Specialized I/O Processors

  - General purpose processors must be able to perform an extremely large number of operations in order to acheive effecient processing. Increasing the number of operations, at a certain point, becomes counter-productive; there are overheads involved in determining the specific operation to be performed, in extra circuitry needed for new operations (which in turn means a physically larger processor with components further apart which means longer delays in signals getting from one area to another and increased power requirements), etc. A processor dedicated to Input/Output does not need a lot of mathematical operations or other advanced functions; all that is needed is the ability to Input or Output, Store to or Load from an indexed memory address, and some simple operations to maintain the "indexing" memory address register. Such an I/O processor therefore may be constructed to be relatively small and fast.

Inter-Processor Communication Options

• The I/O Processor's Instruction Requirements

  Because an I/O processor is a processor in the traditional sense, it needs to have a program stored in some memory area. In general, the I/O processor's program space memory may be shared with the main system processor or it may be unique to the I/O processor (unaccessible to the main processor).

  It should be noted that it is extremely unlikely that the machine level instruction format for the I/O processor will resemble the machine level instruction format for the main system processor. They are (normally) two very different kinds of processors (although there are rare exceptions, as in the case of some LAN-NIC, Local Area Network - Network Interface Cards, which used an Intel 80186 processor for I/O, in IBM-PC type systems whose main processor was a member of the 80x86 family).

• Common Memory Option

  - When the I/O processor's program memory area is available to the main processor, the main processor may use this fact as a means for communicating I/O processing requirements to the I/O processor. This could be done at a very crude level by having the main processor create or modify the actual program instruction codes used by the I/O processor. Alternatively, I/O requirements could be communicated by modifying a parameter table area used by the I/O processor's program.

• Port IO Instruction Option

  - The other method by which the main processor can communicate its I/O processing needs to the I/O processor, is by treating the I/O processor as an Output device to which the main processor sends I/O commands and specifications. The main processor only needs to output a very few bytes of data, in order to direct the I/O processor to perform I/O on what could be thousands or (theoretically) an endless stream of bytes of data.

DMA vs. Port I/O

• Speed

  - An obvious consequence of DMA I/O with a second processor which is performing I/O while the main processor is doing other things, is that processing ellapse-time should be reduced for any task. This assumes equal processors as the "main" processor for each system. While reduced processing time is, in general, the case when using DMA I/O, the speed benefits are not always as great as one might at first think; if the I/O operation involves reading records from a disk file and the main processor can not continue processing until it has those records, the main processor will not be performing anything which advances the completion of task until the I/O processor has completed its job and, thus, the advantage of processing "in parallel" is effectively lost.

• Memory Requirements

  - The memory "mapped" for Input/Output is not available for other uses. This reduces the main processor's "available memory addresses". If the I/O processor's program instruction space is also part of the shared memory, the main processor's available space is even further reduced. The extent to which this is a significant factor depends upon whether the main processor would need to keep a full copy of the I/O physical record anyway (this is often the case).

• Byte vs. Block Requirements

  - Input/Output which can be thought of as occurring as large blocks of transferred data are almost always performed using secondary, I/O processors and DMA. This type of I/O includes disk sector input or output, constant video monitor display, and block data transfer on a LAN or on the InterNet. The main processor is able to do useful work simultaneously with such I/O activities.

  Single byte I/O is more problematic (for example, keyboard input). Here the "set-up" required for a main processor to instruct an I/O processor to perform the input (or output) is virtually equivalent to activity required for the main processor to perform the I/O operation itself.